bertsimas 2011) fairness-efficiency kidney transplantation

8/2/2019 Bertsimas 2011) Fairness-Efficiency Kidney Transplantation

1/47


2/47

Fairness, Efficiency and Flexibility in Organ Allocation for

Kidney Transplantation

Dimitris Bertsimas Vivek F. Farias Nikolaos Trichakis

October 12, 2011

Abstract

We propose a scalable, data-driven method for designing national policies for the allocation

of deceased donor kidneys to patients on a waiting list, in a fair and efficient way. We focus

on policies that have the same form as the one currently used in the U.S. In particular, we

consider policies that are based on a point system, which ranks patients according to some

priority criteria, e.g., waiting time, medical urgency, etc., or a combination thereof. Rather

than making specific assumptions about fairness principles or priority criteria, our method offers

the designer the flexibility to select his desired criteria and fairness constraints from a broad

class of allowable constraints. The method then designs a point system that is based on the

selected priority criteria, and approximately maximizes medical efficiency, i.e., life year gains

from transplant, while simultaneously enforcing selected fairness constraints.

Among the several case studies we present employing our method, one case study designs

a point system that has the same form, uses the same criteria and satisfies the same fairness

constraints as the point system that was recently proposed by U.S. policymakers. In addition,

the point system we design delivers an 8% increase in extra life year gains. We evaluate the

performance of all policies under consideration using the same statistical and simulation tools

and data as the U.S. policymakers use. Other case studies perform a sensitivity analysis (for

instance, demonstrating that the increase in extra life year gains by relaxing certain fairness

constraints can be as high as 30%), and also pursue the design of policies targeted specifically

at remedying criticisms leveled at the recent point system proposed by U.S. policymakers.

MIT Sloan School of Management, [email protected] Sloan School of Management, [email protected] Business School, [email protected]

1


3/47

1. Introduction

Renal or kidney transplantation and maintenance dialysis are the only treatments for end-stage re-

nal disease (ESRD), a terminal disease affecting over 500, 000 people currently in the United States,

see USRDS (2009). Despite b eing a major surgical procedure, transplantation is the treatment of

choice for ESRD patients, as a successful transplantation improves their quality of life. In particular,

dialysis treatment requires that the patient visits a dialysis center for at least 12 hours each week,

whereas transplantation typically allows the patient to resume regular life activities. Furthermore,

a multitude of research and clinical studies have statistically demonstrated that transplantation

also reduces the mortality risk for patients, see Suthanthiran and Strom (1994), Schnuelle et al.

(1998), Port et al. (1993), Ojo et al. (1994). Thus, a kidney transplant is considered by many as a

potentially life-saving gift.

The two sources of kidneys for transplantation are living donors (e.g., family members or friends

of the patient) and deceased or cadaveric donors. The majority of patients are unsuccessful in finding

living donors, and thus join a pool of patients waiting for a deceased donor organ. Of course, while

in the living donor case the donation is typically made to a specific patient, in the deceased donor

case an important allocation problem arises. In particular, once an organ is procured from a

deceased donor, there can be thousands of medically compatible and available recipients the organ

can be allocated to. The problem becomes even more significant, if one accounts for the organ

shortage and the size of the pool of waiting patients in the United States: On October 20th 2010,

86, 391 patients were waiting for a kidney transplant. In 2009, there were 33 , 671 new additions, but

only 16, 829 transplantations were performed, from which 10, 442 transplants were from deceased

donors. For more information and statistical details we refer the reader to UNOS (2010).

In recognition of the aforementioned allocation problem and the growing difficulty of matching

supply and demand, the U.S. Congress passed the National Organ Transplant Act (NOTA) in 1984.

According to this legislation, deceased donor organs are viewed as national resources in the U.S.,

and as such, their allocation has to be based on fair and equitable policies. Moreover, the sale of

organs as well as money transfers of any nature in the acquisition of organs are strictly prohibited.

Instead, the policy for allocating the organs should utilize waiting lists and have the form of a

priority method. That means that patients in need of a transplant register on waiting lists. Then,

2


4/47

once an organ is procured, all medically compatible patients are ranked according to some priority

rules and the organ is successively offered to them according to their ranking, until it is accepted by

a patient. Subsequent to the NOTA, the U.S. Congress established in 1984 the Organ Procurement

and Transplantation Network (OPTN) in order for it to maintain a national registry for organ

matching and develop allocation policies.

Naturally, the aforementioned allocation policies are of central importance and have to accom-

plish major objectives in alleviating human suffering, prolonging life and providing nondiscrimina-

tory, fair and equal access to organs for all patients, independent of their race, age, blood group

or other peculiar physiological characteristics. Some of the main challenges in designing a kidney

allocation policy are the following:

Fairness constraints: What does fair and equal access to organs mean? Due to the subjectivenature of fairness, there is no single fairness criterion that is universally accepted by policy-

makers and academics alike. As such, a great challenge lies in identifying the appropriate

fairness constraints that the allocation outcomes of a policy should ideally satisfy. An exam-

ple of such a constraint could be a lower bound on the percentage of organs allocated to a

particular group of patients say, requiring that at least 47% of all transplants are received by

recipients of blood type O. In the absence of such a constraint these groups would otherwise

be handicapped and not have access to organs because of their physiological characteristics.A number of such criteria have been studied by OPTN policymakers (see OPTNKTC (2008),

RFI (2008)).

Efficiency: As a successful transplantation typically prolongs the life of a patient, while also

improving his quality of life, the policy needs to ensure that the number of quality adjusted

life year gains garnered by transplantation activities is as high as possible. This is also in

line with the view of organs as national resources. Again, this objective is of paramount

importance to the current policy design, see OPTNKTC (2008).

Prioritization criteria: The policy needs to be based on medically justified criteria and phys-

iological characteristics of patients and organs in order to rank patients. However, ethical

rules disallow the use of criteria that can be deemed as discriminatory (e.g., race, gender,

etc.).

3


5/47

Simplicity: Patients need to make important decisions about their treatment options, together

with their physicians. To this end, they need to be able to estimate the probability of receiving

an organ, or at least understand the allocation mechanism. For that reason, the priority

method that is used needs to be simple and easy to communicate.

Implementation: Suppose that one has selected the desired fairness constraints, prioritization

criteria and a simple priority method. How does he then balance the emphasis put on the

different prioritization criteria, so as to design a policy whose allocation outcomes would

maximize efficiency, while satisfying the fairness constraints?

All the above challenges were faced by the OPTN policymakers in 2004, when they initiated

the development of a new national allocation policy that will eventually replace the current one. In

2008, the OPTN released a concrete proposal in a Request for Information publication (RFI (2008))

that is currently under consideration by the U.S. Department of Health and Human Services.

In this work, we deal with the implementation challenge in designing a national allocation

policy, while accounting for all the other challenges above. In particular, we focus on perhaps the

simplest, most common and currently in use priority method, namely a point system. We make

the following contributions:

1. We present a novel method for designing allocation policies based on point systems in a

systematic, data-driven way. Our method offers the flexibility to the policymaker to select

the fairness constraints he desires, as well as the prioritization criteria on which the point

system will be based on. The method then outputs a conforming point system policy that

approximately maximizes medical efficiency, while satisfying the fairness constraints.

2. To validate our proposed method, we utilize it to design policies under different scenarios

of interest to OPTN policymakers. Under a particular scenario, we design a policy that (a)

matches the fairness constraints of the recently proposed policy by U.S. policymakers, and (b)

is based on the same criteria and simple scoring rule format. Critically though, it achieves an

8% increase in anticipated extra life year gains, as demonstrated by our numerical simulations,

which are based on the statistical and simulation tools currently in use by U.S. policymakers

(see below).

4


6/47

3. We use our method to perform a sensitivity analysis that explores the consequences from

relaxing or introducing fairness constraints for instance, what is the impact of reducing the

percentage of transplants to patients on dialysis for greater than 15 years by 1%? In the

case of some constraints, relaxations of fairness constraints can result in life year gains on the

order of 30%. As such, we believe this is a valuable tool in the policy design process.

4. This paper makes a methodological contribution to the area of Approximate Dynamic Pro-

gramming. In particular, we develop a means of designing approximately optimal policies in

problems of dynamic allocation that are massively high dimensional. In particular, these are

allocation problems where the number of classes of objects being allocated, and the num-

ber of bins these objects may b e allocated to are themselves intractably large. A previous

approach to this issue has simply been clustering these classes so as to arrive at a problemwith, say, tens of classes of objects. This approach is a non-starter in the context of designing

real (as opposed to stylized) policies for organ allocation, and we provide a viable alternative

that requires no clustering whatsoever.

Performance in all our numerical studies is evaluated using the same statistical and simulation tools,

as well as data, as the U.S. policymakers use. Those tools and datasets were obtained directly from

their developers, namely the United Network for Organ Sharing (UNOS), which is the non-profit

organization that operates the OPTN, and the Scientific Registry of Transplant Recipients (SRTR).

The structure of this paper is as follows. In the next subsection, we review relevant work in the

literature. Section 2 provides background information on the distribution of organs, the current

allocation policy, as well as updates on the recent development of a new proposed policy. In Section

3, we discuss our method for designing allocation policies in detail. Section 4 includes numerical

evidence of the usefulness of our work through the design of new policies under multiple case

studies, including a sensitivity analysis, and the evaluation of their performance via simulation.

We conclude in Section 5 with an overview of future prospects for the approach delineated in thispaper. A list of acronyms used appears at the end of this paper.

1.1. Literature Review

The model-based analysis of the organ allocation process has attracted significant interest in the

academic literature. One of the first papers in this vain is by Ruth et al. (1985), in which the

5


7/47

authors develop a simulation model to study the problem. Shechter et al. (2005) also introduce

a discrete event simulation model for the evaluation of potential changes to the liver allocation

process. In this work, we utilize the simulation model developed by the SRTR, see KPSAM (2008).

The organ allocation process was also analyzed by Righter (1989) and David and Yechiali (1995)

via a stochastic assignment problem formulation. In their work, they analyze stylized models that

fit into that framework. In this work, we also utilize an assignment problem formulation, but only

for the training phase of our methodology: the output allocation policies of our framework are

rather simple, based on scoring rules and in full compliance with policies that U.S. policymakers

consider, unlike the above referenced work. In a similar vein, Zenios et al. (2000) introduce a fluid

model approximation of the organ allocation process that allows them to explicitly account for

fairness and medical efficiency in the allocation. Our framework accounts for fairness in accordance

with the considerations of policymakers. Zenios (2002), Roth et al. (2004), Segev et al. (2005) and

Ashlagi et al. (2011) study the problem of living donation and the allocation of kidneys. Kong et al.

(2010), Sandikci et al. (2008) and Akan et al. (2011) also tackle the problem of liver allocation.

Another stream of research focuses on the decision-making behavior of patients, by dealing with

organ acceptance policies. David and Yechiali (1985) model the candidates problem as an opti-

mal stopping problem. Similar acceptance policies are developed by Ahn and Hornberger (1996),

Howard (2002), Alagoz (2004) and Alagoz et al. (2007). The present paper will test policies on a

simulator developed by SRTR for OPTN; this simulator assumes a specific, exogenous acceptance

model for patients built from historical data. While the acceptance model ignores endogeneity, it

allows us to simulate outcomes in precisely the manner policy makers currently do.

Recent work by Su and Zenios (2005, 2004) attempts to combine the above streams of research

by explicitly accounting for the acceptance behavior of patients in the development of an allocation

policy. In a similar vein, Su and Zenios (2006) propose an allocation mechanism that elicits the

utilities of the patients. For more details, we refer the reader to the thorough review by Zenios

(2005).

In all the above referenced work dealing with organ allocation policies, the authors design

general near optimal dynamic policies. These papers take the important perspective of designing

a fundamentally new allocation system from the ground up. In our work, we restrict our attention

to policies that comply with the precise constraints imposed by current practice. That is, we focus

6


8/47

our attention on policies based on simple point systems of the precise format as the ones currently

in use and proposed by U.S. policymakers. Moreover, instead of designing a particular policy, we

develop a framework that admits various fairness constraints and prioritization criteria. In other

words, we design a mechanism that can fit directly in the current decision-making process of the

U.S. policymakers.

2. Distribution and Allocation Policies

In this section, we briefly review the distribution process and the operation of the UNOS/OPTN

as coordinators and developers of national policies for the allocation of deceased donor kidneys to

patients. We then discuss the requirements such policies need to meet, and focus on policies that

are based on point systems or scoring rules. Finally, we review the current policy in use in the U.S.

(which itself is based on a scoring rule), as well as updates on the development of a new scoring

rule based national policy.

In the U.S., the non-profit Organ Procurement Organizations (OPOs) are directly responsible

for evaluating, procuring and allocating donated organs within their respective designated service

area. Once consent is obtained and an organ is procured by an OPO, the OPTN computerized

national registry automatically generates a list of patients, who are medically compatible with the

procured organ. Medical compatibility of patients is determined by their physiological characteris-

tics and those of the procured organ (e.g., accounting for ABO incompatibility1, weight and size,

unacceptable antigens, etc.). Subsequently, the priority method used by the OPO determines the

order in which the organ will be offered to patients. Once a kidney is procured, it can be typically

preserved for up to 36-48 hours, after which the organ can no longer be used for transplantation.

For that reason, priority is given to local patients, although there are rules that determine when

priority should be given to non local patients. After an offer is b eing made to a patient, he has

to decide with his surgeon whether to accept it or not within a limited amount of time. In case

of rejection, the organ is offered to the next patient according to the specified order and so on. In

case no patient accepts the organ within 36-48 hours, the organ is discarded.

In addition to using the OPTN national registry, the activities of the OPOs, and their allocation

1ABO incompatibility is a reaction of the immune system that occurs if two different and not compatible bloodtypes are mixed together, see http://www.nlm.nih.gov/medlineplus/ency/article/001306.htm.

7
http://www.nlm.nih.gov/medlineplus/ency/article/001306.htmhttp://www.nlm.nih.gov/medlineplus/ency/article/001306.htm


9/47

policies in particular, are coordinated and regulated by the OPTN. That is, the OPTN provides

general guidelines and lays out a national allocation policy that is suggested to all OPOs. The

allocation policy of every OPO then needs to be consistent with the national policy, although

minor alterations are possible subject to approval by the OPTN.

2.1. National Allocation Policies

National policies for the allocation of the deceased donor kidneys are developed by the OPTN

Kidney Transplantation Committee (KTC), and are approved by the U.S. Department of Health &

Human Services. Policies need to account for numerous legal, economic, institutional, ethical, and

other societal factors; the requirements for an allocation policy are included in the OPTN Final

Rule (DHHS (2000)). Below we summarize the most important guidelines that policies have to

conform to as per the OPTN Final Rule. In particular, the allocation

(a) shall seek to achieve the best use of donated organs, and avoid organ wastage;

(b) shall set priority rankings based on sound medical judgment;

(c) shall balance medical efficiency (extra life years) and equity (waiting time), without discrim-

inating patients based on their race, age and blood type;

(d) shall be reviewed periodically and revised as appropriate.

Additionally, the priority method in place needs to be simple and easy to communicate, as

discussed in the Introduction. As such, the ranking of patients is typically achieved by means of a

point system or scoring rule: all national allocation policies that have been used in practice have

been based on scoring rules. We formally define next the notion of a scoring rule based policy and

then discuss the current national policy and suggested revisions.

Point system or Scoring rule based policies. Under a policy based on a scoring rule, patients

are ranked according to a calculated score, commonly referred to in this context as the Kidney

Allocation Score (KAS). Specifically, a scoring rule consists ofscore components and scalar constant

score weights. A score component can be any function of the characteristics of a patient and/or an

organ. Then, once an organ is procured and needs to be allocated, one calculates the individual

score components for each patient and the particular procured organ. The KAS for each patient

8


10/47

is evaluated as the weighted sum of his score components (using the score weights). To introduce

some notation, given a patient p and an organ o, we denote the jth score component with fj,(p,o),

and the jth score weight with wj . The KAS of patient p for receiving organ o, KAS(p,o), is then

calculated as

KAS(p,o) =j

wjfj,(p,o).

For instance, examples of score components can be the number of years the patient has b een

registered on the waiting list for, the life expectancy of the patient in case he remained on dialysis,

or the life expectancy in case he received the procured organ, etc.

One can think of a scoring rule based policy as a priority method that awards points to patients

based on different criteria (the score components); patients are also potentially awarded different

amounts of p oints per criterion, based on the score weights. The ranking is then achieved based

on the number of points collected by each patient. The current policy in use and the one recently

proposed by U.S. policymakers are both examples of scoring rule based policies and are discussed

next.

Current allocation policy. The current policy has been in existence for more than 20 years.

It is based on a scoring rule that utilizes waiting time, a measure of the patients sensitization 2

and tissue matching3 of the organ and the patient as score components. The rationale behind

this rule is as follows. Points are given for waiting time and sensitization in order to serve the

fairness objective of the allocation and to provide equal access to organs to all patients (note that

highly sensitized patients have reduced medical compatibility with donors). On the other hand,

since tissue matching is an indication for a successful transplantation, the points given to matched

patients serve the medical efficiency objective of the allocation. For more details we refer the reader

to ODADK (2010).

Recent advances in medicine and changes in patients needs however, have rendered the current

policy inappropriate. More specifically, these changes have rendered the current policy inconsistent2Potential recipients are sensitized if their immune system makes antibodies against potential donors. Sensitiza-

tion usually occurs as a consequence of pregnancy, blood transfusions, or previous transplantation. Highly sensitizedpatients are more likely to reject an organ transplant than are unsensitized patients. For more information, seehttp://www.ustransplant.org/

3When two people share the same human leukocyte antigens (abbreviated as HLA), they are said to be amatch, that is, their tissues are immunologically compatible with each other. HLA are proteins that arelocated on the surface of the white blood cells and other tissues in the body. For more information, seehttp://www.stanford.edu/dept/HPS/transplant/html/hla.html

9
http://www.ustransplant.org/http://www.stanford.edu/dept/HPS/transplant/html/hla.htmlhttp://www.stanford.edu/dept/HPS/transplant/html/hla.htmlhttp://www.ustransplant.org/


11/47

with the OPTN Final Rule, see Norman (2009) and RFI (2008). For instance, the long waiting

times experienced by the patients, coupled with advances in medicine that have prolonged the

survivability of patients on dialysis, have resulted in the accumulation of points for waiting time

by the patients. This accumulation of points has then created an imbalance between the efficiency

and fairness objectives of the allocation, see OPTNKTC (2007). In response to that, and in line

with the requirement of the OPTN Final Rule for periodic review of the policy, the KTC has been

reviewing the policy for the past few years and is currently in the process of developing a new

policy, see OPTNKTC (2007).

Development of a new policy. Since 2004, the KTC has considered more than 40 different

scoring rules, which involve various score components, see OPTNKTC (2010). We first discuss

the criteria upon which the score components are based, and then discuss the components. For a

patient p and an organ o, the criteria are

1. Tissue matching or HLA matches, i.e., the number of HLA shared by patient p and organ o;

2. Age of patient p and/or donor of organ o, denoted by AGE(p) and AGE(o);

3. Wait time, which is equal to the number of years patient p has been registered at the waitlist;

4. Dialysis time, which is equal to the years patient p has spent on dialysis, denoted by DT(p);

5. Blood group of patient and/or donor;

6. Expected post transplant survival of patient p from receiving organ o;

7. Expected waitlist survival of patient;

8. Life years from transplant, denoted by LYFT(p,o), which is equal to the expected incremental

quality-adjusted life years gain of patient p from receiving organ o, compared to remaining

on dialysis (for a precise definition, we refer the reader to Wolfe et al. (2008));

9. Donor profile index, denoted by DPI(o), which is a number between 0 and 1, indicating the

quality of the donated organ (0 corresponds to an organ of highest quality);

10. Calculated panel reactive antibody, denoted by CPRA(p), which is a number between 0 and

100, measuring the sensitization of the patient (0 corresponds to the lowest level).

10


12/47

A typical scoring rule proposed by the KTC includes 3-5 score components that are functions of

(some of) the above criteria. In most cases, the components are either linear functions (e.g., points

are awarded per year on dialysis, or per life year from transplant, etc.), or nonlinear functions

of one or more criteria (e.g., for patient p and organ o, points are awarded according to (1

DPI(o)) LYFT(p, o), or DPI(o) DT(p), or |AGE(p) AGE(o)|, etc.), including stepwise or

indicator functions (e.g., points are awarded if patient p is highly sensitized, CPRA(p) 80, or if

he is aged less than 18 or 35, AGE(p) 18 or 35, etc). For more details, we refer the reader to

OPTNKTC (2007), OPTNKTC (2008).

As mentioned above, the KTC considered more than 40 different scoring rules, each of which

utilizes a subset of the score components above. Furthermore, based on simulation experiments,

the KTC evaluated the performance of the proposed scoring rules and identified weights that were

deemed appropriate (see OPTNKTC (2008)). The dominant proposal up to this point, published in

2008 in a Request For Information document (RFI (2008)), entails the following score components:

LYFT (1 DPI), DT DPI, DT and CPRA. The associated score weights are 0.8, 0.8, 0.2 and

0.04. That is, the Kidney Allocation Score under the dominant proposal is

KAS(p,o) = 0.8 LYFT(p,o) (1 DPI(o))

+ 0.8DT(p) DPI(o)

+ 0.2DT(p)

+ 0.04 CPRA(p).

The first two components are the life years from transplant and dialysis time, scaled by the donor

profile index. The scaling ensures that in case of a high quality organ (DPI close to 0), emphasis

is given on life years from transplant, whereas in case of a low quality organ (DPI close to 1),

emphasis is given on dialysis time. The last two components are the dialysis time and calculated

panel reactive antibody score of the patient. More information and motivating aspects can be found

within the Request For Information document (RFI (2008)). As an example, consider an organ o of

medium quality, with DPI(o) = 0.55. Then, patients are awarded 0.8 (1 0.55) = 0.36 points for

every quality adjusted incremental life year they would gain in expectation, 0 .8 0.5 5 + 0.2 = 0.64

points for every year they have spent on dialysis, and 0.04 points for every point of their CPRA

11


13/47

score.

While medical expertise and the OPTN Final Rule can guide the identification of the score

components of clinical validity, the task of finding the right selection or subset of these components

and the appropriate weights is more involved, as the experimentation of the OPTN KTC with

more than 40 different rules suggests. A natural question in response to the proposed scoring rule

is whether this is the best we can do. In particular, does there exist another scoring rule of the same

format, based on the criteria and score components considered by policymakers, that dominates

the proposed one, i.e., is equally or more fair and more efficient? Admittedly, this is an involved

question to answer; to illustrate this, consider only changing the weights in the proposed scoring rule

above. The outcomes by such a change can p erhaps be evaluated only via simulation; simulating

a single specific scoring rule takes hours. This severely curtails the efficacy of a search for a policy

that while p ossessing the requisite fairness properties is also efficient. Our proposed methodology

provides a valuable tool in this search and takes a step towards answering the questions posed

above.

3. Designing Allocation Policies

We propose a method for designing scoring rule based policies for the allocation of deceased donor

kidneys to patients. Specifically, we propose a data-driven method that computes in a systematicway score weights associated to pre-specified score components, so that the resulting policy achieves

a near-optimal medical utility (measured by life years from transplant gains). In other words,

after one has decided upon the components he wishes to include in a scoring rule, our method

utilizes historical data to efficiently compute associated weights, so as to maximize the efficiency

of the policy. In addition, our method can also take as input fairness constraints on the allocation

outcomes; while we defer the precise definition of the class of admissible constraints for Section 3.1,

we note here that our method captures a multitude of important and commonly studied constraintsof interest to policymakers. Then, the method computes the score weights, so that the resulting

policy is as efficient as possible, and the fairness constraints are approximately satisfied.

Figure 1 illustrates the functionality of the proposed method. Typically, policymakers select

their desired score components that would feature in the scoring rule and constraints that the

allocation outcomes need to satisfy. Our method provides an efficient, scalable and systematic

12


14/47

score weights

historical data

score components

fairness constraints

Figure 1: An illustration of the functionality of the proposed method.

way of striking the right balance between the selected score components by designing a policy that

approximately maximizes medical efficiency, subject to the selected constraints.

As an application of our method, we use historical data from 2008, to construct multiple scoring

rule based policies that utilize the same criteria for components as the ones considered by the OPTN

Kidney Transplantation Committee. Within the different case studies we present, we also design a

policy that possesses similar fairness characteristics with the KTC dominant proposal. Numerical

studies then suggest that this policy constructed by our method achieves an 8% improvement

in life years from transplant, using the same statistical and simulation tools and data as U.S.

policymakers use. Furthermore, we perform a trade-off analysis by considering deviations from the

fairness constraints of the proposed policy. In particular, we study the effect in life year gains of

the policy, in case of emphasizing or deemphasizing the priority given to patients who have beenwaiting for a long time or are sensitized. Our method efficiently redesigns policies accordingly.

The results indicate that the performance gain in life years from transplant can be as high as 30%.

Details on the case and simulation studies are included in Section 4.

We next present our proposal in full detail.

3.1. Methodology

Given a list of n score components, related historical data of patients and donated organs charac-teristics, and constraints on the allocation outcomes (precisely defined below), we calculate score

weights w1, . . . , wn, such that the resulting scoring rule policy satisfies the constraints approxi-

mately, while maximizing life years from transplant.

Consider a fixed time period over which we have complete (ex facto) information about all

patients registered in the waitlist (pre-existing and arriving) in that time period. In particular,

13


15/47

we know their physiological characteristics, the time of their initial registration, as well as the

evolution of their medical status and availability for a transplant during that time period. Suppose

we also have complete information about the organs that are procured during the period, that is

the time at which they are procured and their physiological characteristics. We index the patients

by p = 1, . . . , P and the organs by o = 1, . . . , O. We say that patient p is eligible to receive organ

o, or equivalently that the patient-organ pair (p,o) is eligible for transplantation, if at the time of

the organ procurement all conditions below are met:

1. The patient is registered at the waitlist for a transplant;

2. The patient is actively waiting for a transplant and his medical status is appropriate for

transplantation;

3. The patient is medically compatible with the organ.

Let C be the set of patient-organ pairs eligible for transplantation, i.e.,

C = {(p,o) : patient p is eligible to receive o} .

Note that one can construct C simply by using the arrival information and characteristics of the

organs and the patients, and the evolution of the availability and medical status of the patients.

Additionally, one can also compute the score components for each eligible patient-organ pair,

as well as the life years from transplant. Let fj,(p,o) be the value of the jth component score,

j = 1, . . . , n, and LYFT(p,o) the life years from transplant for pair (p,o) C.

We now define the class of admissible constraints on the allocation outcomes, alluded to thus

far. First, let x(p,o) be defined for every eligible patient-organ pair (p,o) as

x(p,o) =

1, if organ o is assigned to patient p,

0, otherwise.

A constraint is admissible for our method if it is linear, that is if it can be modeled as a linear

constraint with respect to variable x. The class of constraints that can be modeled in this way is

very broad, and captures the majority of constraints a policymaker might wish to incorporate; for

instance, one can impose lower bounds for a specific group of patients on

14


16/47

the probability of receiving a transplant,

the average life years from transplant gained among the actual transplant recipients,

the average time spent on dialysis among the actual transplant recipients.

As an example, a lower bound L on the number of organs allocated to a specific group of patients

G {1, . . . , P }, can be expressed as

pG

o:(p,o)C

x(p,o) L.

For instance, setting G to be the set of all patients of blood type O could enforce a lower bound on

transplants for patients of this blood type.

We denote the input fairness constraints with Ax b, for some matrix A and vector b. We now

present our method which consists of three steps:

Step 1 (An Idealized Matching Problem): Consider a social planner with foresight who has

knowledge of the set of all eligible pairs C and the life years from transplant score for every pair in

the set. Suppose also that patients accept all organs offered to them. In this setup, the problem of

allocating organs to patients so as to maximize medical efficiency, i.e., life years from transplant,

subject to fairness constraints Ax b, can be formulated as a linear optimization problem:

maximize

(p,o)C

LYFT(p,o)x(p,o)

subject to

o:(p,o)C

x(p,o) 1, p

p:(p,o)C

x(p,o) 1, o

Ax b

x 0.

(1)

Note that a fractional value for x(p,o) can be interpreted as the probability of assigning organ o

to patient p in a randomized policy. Its solution suggests an allocation with perfect hindsight

(as opposed to an implementable policy). The next two steps will use this idealized solution to

construct an implementable policy in a unique way.

Step 2 (Dual Information): By linear optimization duality, if y is the vector of optimal dual

15


17/47

multipliers associated with the constraints Ax b for problem (1), then problem (1) is equivalent

with the one below:

maximize

(p,o)C

LYFT(p,o)x(p,o) yTAx + yTb

subject to

o:(p,o)C

x(p,o) 1, p

p:(p,o)C

x(p,o) 1, o

x 0.

(2)

Note that problem (2) is a matching problem. We equivalently rewrite the objective of (2) as

cTx + yTb, utilizing the cost vector c defined as

c(p,o) = LYFT(p,o)

yTA(p,o)

, (p,o) C.

We next use this dual information to construct an implementable policy.

Step 3 (Approximate Dynamic Programming): Note that our goal is to design a policy that

approximately solves the above matching problem online, i.e., a policy that sequentially matches

organs at their time of procurement to available patients without utilizing any future information.

An implementable policy will require the following:

1. An estimate of the value of assigning a particular organ, o to a particular patient, p (tech-

nically, one may think of this as a differential value function for the associated stochastic

optimization problem).

2. An interpretable formula for the above differential value that uses permissible features of the

patient and organ in a clinically acceptable way. Our goal is to rank patients not by any

artificial score coefficients, but rather based on the selected score components.

One possible policy is scoring potential allocations on the basis of the coefficients c(p,o) computed

above. Unfortunately, the c(p,o) coefficients are calculated for patients on the waiting list and

received organs from some historical data set and as such it is likely that we will not have access

to c(p,o) for all pairs (p,o) moving forward. More importantly, a scoring policy based on these

coefficients will not satisfy the second requirement above. As such, we consider using the coefficients

16


18/47

to inform the calibration of an acceptable scoring rule. In particular we find acceptable score weights

w by solving the optimization problem

minimize (p,o)C

c(p,o) w0 n

j=1

wjfj,(p,o)

2

subject to w S,

where the set S enforces clinical and ethical requirements (for instance, by requiring that the

resulting policy be continuous or monotone in certain score components, etc.).

The method is summarized as Procedure 1.

Procedure 1 Computation of score weights

Input: list of n score components, data for linear constraints (A, b), historical data: set of eligible

patient-organ pairs C, life years for transplant LYFT(p,o) and values of score components,fj,(p,o), j = 1, . . . , n, for every eligible pair (p,o).Output: weights for scoring rule, w1, . . . , wn.

1: solve problem (1)2: y vector of optimal dual multipliers associated with constraints Ax b

3: c(p,o) c(p,o) = LYFT(p,o)

yTA(p,o)

, (p,o) C

4: use (potentially constrained) linear regression to find w0, w1, . . . , wn, such that for all (p,o) C

c(p,o) w0 + w1f1,(p,o) + . . . + wnfn,(p,o).

3.2. Discussion

In this section, we discuss (a) why one should expect the proposed method to perform well in

practice, and (b) the relative merits of our contribution.

Consider the airline network revenue management setting analyzed in Talluri and van Ryzin

(1998). In that setting, an airline is operating flights and is selling different itinerary tickets

to incoming customers, so as to maximize net expected profits from sales subject to capacity

constraints (which correspond to the numbers of seats on the different aircrafts operating the

flights). The authors analyze a simple control policy that decides whether to sell an itinerary

ticket to a passenger or not, and demonstrate that the policy is asymptotically optimal under some

conditions. For the organ allocation problem, a simplified version of the policy that we described

in the previous section can be cast in the same framework as in Talluri and van Ryzin (1998);

17


19/47


20/47

3. From a methodological p erspective our approach can b e seen as a contribution to approxi-

mate dynamic programming techniques in general and bid-pricing methodologies for dynamic

allocation in particular. More precisely, bid pricing methodologies exist for the dynamic allo-

cation of a potentially large set of objects arriving randomly over time and each of which can

be classified into a smallnumber of distinct types, to a small set of bins, each with potentially

large capacity. In the kidney allocation context the heterogeneity in patients and organs pre-

cludes classifying either into a small number of types. Our methodology shows how to deal

with this problem and in general extends bid price methodologies to scenarios where arriving

objects and bins are highly heterogeneous.

4. The failure of the current kidney allocation policy in place to keep up with advances in

medicine and the changes in patients needs throughout the years, has demonstrated that insuch a dynamic and complex environment, revisions to policies are likely to be required in the

future as well, a fact that is also recognized by the OPTN Final Rule. Furthermore, even in

the current process of developing a new policy, there is no guarantee that the Office of Civil

Rights will approve the criteria of life years from transplant, dialysis time, etc., suggested by

the OPTN policymakers. In both cases, our method will expedite the development of a new

policy, as it would require only an updated list of score components and fairness properties

to be specified.

5. Our method allows for sensitivity analysis; specifically, one can efficiently evaluate the out-

comes of relaxing some or introducing new fairness constraints. In the next section, we provide

such an analysis that reveals the dependence of medical efficiency on fairness concepts, and

illustrate how it can be used in practice by policymakers. In particular, note that one of the

main goals that the OPTN policymakers have set for a new national policy is to deempha-

size the role of waiting time and increase medical efficiency (see Section 2.1). Our analysis

provides a characterization of the trade-offs involved.

In the next section, we provide numerical evidence of the usefulness of the described method. In

particular, we use historical data to design multiple new scoring policies under different scenarios

and also perform a sensitivity analysis.

19


21/47

4. Numerical Evidence and the Design of New Allocation Policies

We utilize the method described in the previous section to design new scoring-rule based policies

for kidney allocation that have different fairness requirements and/or are based on different score

components. Specifically, we consider various fairness requirements and score components derived

from policies that have been proposed by the OPTN Kidney Transplantation Committee (see

Section 2.1), to set up 4 realistic case studies. Briefly, the intent and outcomes of these case studies

are as follows:

1. Case Study 1: Here we demonstrate the utility of our approach in finding policies using

precisely the set of score components utilized by the current dominant proposal considered

by the KTC (referred to also as the KTC p olicy or proposal in this section). Using the

methodology of the previous section we require the approach to preserve the fairness properties

of the dominant proposal. In addition to finding a policy that is de facto clinically valid in an

automated fashion, the policy designed provides life year gains of approximately 3% relative

to the dominant proposal, in spite of being so heavily constrained. This demonstrates the

value of the approach in designing policies given requirements on fairness, and a small set of

score-components.

2. Case Study 2: Here we allow policies that can potentially be based on all score components

considered by policy makers but continue to require the fairness properties of the dominant

proposal. As described in the previous Section, we impose constraints on our methodology

to guarantee that the resulting scoring rule is clinically valid. Our methodology produces a

policy that in addition to being clinically valid and exhibiting similar fairness properties as

the dominant proposal provides an 8% increase in life year gains relative to that proposal. In

addition to the merits exhibited via the first case study, this demonstrates the value of the

approach vis-a-vis the task of guiding the selection of a small but appropriate set of score

components from a large family of potential score components.

3. Case Studies 3 and 4: Similar to the previous study, this study requires fairness criteria

that correspond to a perhaps more balanced allocation of organs among different age groups

of patients (as opposed to the dominant proposal which has been criticized as providing too

few transplants to older age groups). Surprisingly, we develop a clinically acceptable p olicy

20


22/47

that allays these criticisms of the dominant proposal while providing essentially the same

life years gains as the dominant proposal. This demonstrates the value of our methodology

vis-a-vis designing policies that must satisfy stringently specified fairness criteria. We further

consider various relaxations of the fairness criteria and construct corresponding policies to

show how our approach allows policy makers to perform a sensitivity analysis relative to

fairness requirements.

To ensure a fair comparison, we evaluate the performance of all policies we study by using the

same statistical models and tools, as well as datasets with the OPTN KTC policymakers. We first

provide details about the data and models, and then present our methodology and results.

4.1. Data, Statistical Models and Tools

This work uses highly detailed historical data from the Scientific Registry of Transplant Recipients

(SRTR). The SRTR data system includes data on all donor, wait-listed candidates, and transplant

recipients in the U.S., submitted by the members of the Organ Procurement and Transplantation

Network (OPTN). The datasets include all the various physiological and demographic characteris-

tics of wait-listed patients and donors that are needed for our study, as well as the evolution of the

medical status of the patients, and the arrival process of the donated organs.

In addition, the SRTR has developed sophisticated survivability models for ESRD patientsusing historical survival rates. The models provide an estimate for the anticipated lifespan of a

patient in case he remained on dialysis, or in case he received a particular kidney, based on a

plethora of physiological attributes (e.g., the patients age, body mass index, diagnosis, as well

as tissue matching, the donors age, cause of death, etc.). For more information and a detailed

study of the statistical performance of the models, we refer the reader to Wolfe et al. (2008) and

Wolfe et al. (2009). The SRTR has also developed an acceptance model that predicts the probability

of a particular patient accepting a particular organ offered to him, based on the physiologicalcharacteristics of the patient and the donor, the distance, etc.

The above datasets and statistical models have also been utilized by the SRTR in the devel-

opment of the Kidney-Pancreas Simulated Allocation Model (KPSAM). The KPSAM is an event-

driven simulator that simulates the entire allocation process using historical data, for different

allocation policies. It was developed in order to support studies of alternative policies. The KP-

21


23/47

SAM is the platform that the OPTN KTC is utilizing to evaluate the performance of their proposed

policies, see OPTNKTC (2007). For more details on the data and the simulator, we refer the reader

to Waisanen et al. (2004) and KPSAM (2008).

For the purposes of this study, we obtained the KPSAM and utilized its simulation engine

in order to obtain realistic allocation outcomes of the policies we consider. The life years from

transplant gains are estimated using the aforementioned survivability models, embedded in the

KPSAM.

4.2. Methodology

We perform 4 case studies of designing scoring rules that have different fairness requirements

and/or are based on different score components. In all studies we design allocation policies using

our method described in Section 3. Recall that our method outputs scoring rules, given as input

historical data, fairness requirements and score components. We discuss the specific fairness re-

quirements and score components of each case study separately (see below). For historical data, we

use the first 6 months of data of the 2008 SRTR database as input to our method (training data).

To evaluate the performance of a policy, we use the KPSAM to simulate the allocation outcomes

of that policy 100 times, over the remaining 6 months of the 2008 dataset that were not used as

training data. To evaluate efficiency, we record the average number of transplantations occurring

and the average net expected life years from transplant (along with sample standard deviation).

To compare fairness properties across different policies, we compare the percentage distribution of

transplant recipients across different races, age groups, blood types, sensitization groups, as well as

diagnosis types and years spent on dialysis. Note that this practice is in line with the comparison

criteria studied by the OPTN policymakers (see OPTNKTC (2008), RFI (2008)). As such, we also

record the average aforementioned percentage distributions for the 100 simulation runs (along with

sample standard deviations).

We next present the case studies and discuss our results. In the first study, we design a policy

that utilizes the exact same score components as the dominant KTC policy, while exhibiting the

same fairness properties. Similarly, we also design a policy that is based on the exact same score

components as another policy considered by the KTC. In the second and third case studies, we

consider all score components and criteria studied by the KTC (see Section 2.1), and actually use

22


24/47

our method to select a subset of them, while enforcing the same fairness properties as the KTC

policy (Case Study 2), or alternative fairness properties (Case Study 3). The fourth case study

deals with a trade-off analysis, where we further study deviations from the fairness properties of

the KTC proposed policy and their consequenses to efficiency.

4.3. Case Study 1

We design two policies that are based on the same score components as two specific policies proposed

by KTC, including the dominant proposal. We enforce the policies to have the same fairness

properties as the dominant KTC proposal.

Score components. We use (a) the same score components as in the KTC policy, namely (1

DPI) LYFT, DPI DT, DT, CPRA, and (b) the same score components as in another KTC

proposed policy5, namely (1 DPI2) LYFT, DPI2 DT, DT, CPRA.

Fairness constraints. We require the two policies to have the same fairness properties as the

dominant KTC proposal. To enforce that, we simulate the KTC policy and record the percentage

distribution of transplant recipients across the different groups discussed above. We then use those

recorded percentage distributions as input constraints. More specifically we use them to input

lower bound constraints on the percentage of organs allocated to the following groups: Caucasian,

African-American, Hispanic or patients of another race; patients aged between 18-35, 35-50, 50-

65 and above 65 years; patients who have spent less than 5, 5-10, 10-15 or more than 15 years

on dialysis; blood type O, A, B, AB patients; patients diagnosed with nephritis, hypertension,

polycystic kidney disease, diabetes or other disease; patients with a sensitization level (CPRA) of

0-10, 10-80 or 80-100. For instance, consider the fairness constraints pertaining to dialysis time.

The recorded p ercentage distribution of recipients for the KTC policy is as follows: 54% of the

recipients have spent less than 5 years on dialysis, 29 .5% between 5-10 years, 11.1% between 10-15

5One of the policies considered by the KTC entailed an allocation score similar to the score of the dominant KTC

policy, but with the DPI term replaced by DPI

2

.

23


25/47


26/47


27/47

percentage of transplant recipients

65 yr

nephritis

hypertension

polycystic

other

diabetes

CPRA 0-10

CPRA 10-80

CPRA 80-100

0 20 40 60

Figure 3: Simulated average percentage distributions of recipients across different sensitization, diag-nosis and age patient groups for the benchmark KTC policy (blue) and the policies designed in Sections4.3-4.4, i.e., in Case Study 1(a) (cyan), in Study 1(b) (yellow) and in Study 2 (red). The results are foran out-of-sample period of 6 months in 2008.

26


28/47

This case then demonstrates that our method reliably designs policies that perform well based on

input that is directly related to outcomes (fairness constraints) and permissible score components.

In contrast, given a set of score components, the approach of policymakers in designing policies has

traditionally been to first identify weights and then observe (simulation based) outcomes; and then

go through those steps multiple times if necessary. This is obviously not ideal; a fact demonstrated

amply by policy (a) designed via our methodology above. Thus, our contribution enables the design

of policies in a more natural and powerful way by considering the desired outcomes directly. In the

next case study we show how our method can be used to select appropriate score components as

well.

4.4. Case Study 2

In this case study, we design a policy that has the same fairness properties as the dominant KTC

policy and is based on criteria and score components considered by the KTC. Specifically, we allow

the policy to use any (small) subset of those criteria and components. In addition by imposing

the appropriate constraints in the third (regression) phase of our methodology we ensure that

the resulting policies are clinically valid (i.e., they conform qualitatively to features of past KTC

policies/ recommendations).

Score components. Rather than pre-selecting specific score components, we feed the regression

step of our method with virtually all score components considered by the KTC (see Section 2.1).

However, in accordance with the format of KTC policies, we eventually pick only the 4 most

significant components, which prove to be: LYFT, a continuous piece-wise linear function of DT

(with potential break points at 5 and 10 years), CPRA and a step-wise function of patient age

(with potential break points at 50 and 65 years) that gives additional points for passing 50 years of

age, and 65 years of age. These last two score components are in line with KTC proposals, and in

fact, make the policy highly desirable by addressing certain ethical issues raised over equity across

age groups as we discuss later.

Fairness constraints. As in Case Study 1, we enforce the policy to have the same fairness

properties as the dominant KTC policy.

Results. The output of our method is the scoring rule assigning the Kidney Allocation Score to a

27


29/47

KTC policy Case Study 2 Case Study 3

number of transplantations (std) 5, 799 (23) 5, 807 (22) 5, 822 (24)net life years from transplant (std) 34, 217 (185) 36, 890 (219) 34, 065 (212)

Table 2: Simulation results of the KTC policy and the policies designed in Case Studies 2 and 3 inSections 4.4-4.5, for an out-of-sample period of 6 months in 2008 and 100 runs.

patient-organ pair (p,o) of

KAS(p,o) = LYFT(p,o) + g (DT(p)) + 0.08 CPRA(p) + 0.5 I (AGE(p) 50) ,

where I is the indicator function and

g (DT) =

0.65DT, 0 DT 5,

DT 1.75, 5 DT 10,

0.2 DT + 6.25, 10 DT.

According to the above scoring rule, patients are awarded 1 point for every life year from transplant

gain, 0.08 points per point of their sensitization score, 0.5 points if aged more than 50 and points

based on their dialysis time as follows: 0.65 points for the first 5 years, 1 point for every additional

year up to 10 years and 0.2 points for every additional year beyond that.

The simulation results are reported in Table 2 and Figures 2, 3. Sample standard deviations

for the percentage distributions are included in the Appendix.

Discussion. In this case study we attempt to address the question: given all the score components

and criteria we can use and the fairness properties of the KTC policy, can we design a policy

that is based on some of those components, has the same fairness properties but is more efficient?

The results demonstrate that our method is indeed capable of doing so, as the policy we design

delivers an 7.8% increase in life year gains in comparison to the KTC policy. The score components

our policy uses are all based on components and criteria the KTC has already considered and are

discussed next.

The designed policy awards points according to life years from transplant (LYFT), dialysis

time (DT), sensitization level (CPRA). The policy also uses a step-wise score component based

on patient age that has the same form as the component pertaining to sensitization in the current

28


30/47

policy in use, see UNOS (2010). Note also that the policy uses patient age in a manner that

allays critiques of earlier KTC proposals. In particular, age has b een used by the KTC primarily

to direct more or higher quality organs to younger patients for efficiency purposes. For instance,

the allocation policy currently in use in the U.S. gives priority to pediatric patients (aged less

than 18 years) for organs procured from donors aged less than 35, whereas proposals suggest to

extend priority to patients aged less than 35 as well. This may be perceived as providing an undue

advantage to younger patients. In contrast, the way our policy utilizes age is in the other direction.

That is, to impose fairness, the policy awards points to patients aged more than 50 to compensate

for the fact that they typically obtain smaller LYFT scores.

Futhermore, note that the score component pertaining to DT is continuous piece-wise linear.

We present here the use of a piece-wise linear function for the following reasons: (a) to illustrate

how our method can deal with a variety of functions for score components, and (b) the preference of

some members of the KTC for continuous functions6, due to the concern that discontinuities might

grant patients who are slightly above a particular threshold (or breakpoint) a disproportionate

advantage compared with patients slightly below that threshold, see OPTNKTC (2007). Note

however, we found that in all policies presented in this work, one can interchangeably use step-wise

and piece-wise functions (delivering statistically indistinguishable performance).

Finally, this case study demonstrates how our method is capable of calibrating multiple score

components simultaneously and distilling the ones that are important. In the absence of an algo-

rithmic method, such a task might be strenuous or even impossible to carry out.

4.5. Case Study 3

We present a case study similar to Case Study 2, but with different fairness requirements. We

enforce the same fairness properties with the KTC policy, but require the percentages of recipients

aged 50-65 and above 65 to be equal to the percentages of patients at the waitlist aged 50-65 and

above 65 respectively. This requirement results in an (additive) increase of around 17 p ercentage

points for organs allocated to patients aged 50 and above. This increase is balanced by a pro-rata

decrease in the number of organs allocated to recipients aged between 18 and 50. We require

this at the outset in response to comments made by a UNOS Ethics Committee in OPTNKTC

6However, note that step-wise score components are utilized in the current allocation policy.

29


31/47

(2008) that observed that the dominant KTC proposal resulted in a decrease in the proportion of

transplants among patients in the 50-65 and 65 and older age groups. The question that remains

is whether the life year gains provided by the dominant proposal can be retained using a clinically

acceptable policy that unlike the dominant proposal, does notresult in a large change in the fraction

of transplants among older age groups.

Score components. As in Case Study 2, we use LYFT, a piece-wise linear function of DT (with

break points at 5 and 10 years), CPRA and a step-wise function of patient age (with break points

at 50 and 65 years).

Fairness constraints. We enforce the policy to have the same fairness properties as the domi-

nant KTC policy, but with a different age distribution requirement. In particular, the p ercentage

distribution of patients according to age for the KTC policy is: 5% for patients aged less than 18,

20.4% for patients aged 18-35, 32.4% for patients aged 35-50, 30.7% for patients aged 50-65 and

11.5% for patients aged above 65. For our policy, we require the percentages of organs allocated to

patients aged 50-65 and above 65 to be 41% and 17.8% respectively (note that those were preci-

cely the percentages of patients aged 50-65 and above 65 in the waitlist in 2008). Accordingly, we

require the percentages of organs allocated to patients aged 18-35 and 35-50 to be 14% and 22 .2%

respectively.

Results. The output of our method is the scoring rule assigning the Kidney Allocation Score to a

patient-organ pair (p,o) of

KAS(p,o) = LYFT(p,o) + h (DT(p)) + 0.12 CPRA(p) + 2.5 I (AGE(p) 50) + I (AGE(p) 65) ,

where

h (DT) =

0.75DT, 0 DT 5,

DT 1.25, 5 DT 10,

0.5 DT + 3

.75

,10 DT

.

The simulation results are reported in Table 2 and Figure 4. Sample standard deviations for the

percentage distributions are included in the Appendix.

Discussion. This case study, alongside the next one, illustrates how our method deals with

alternative fairness constraints. In particular, we consider the same setup as in Case Study 2,

30


32/47

percentage of transplant recipients

65 yr

nephritis

hypertension

polycystic

other

diabetes

CPRA 0-10

CPRA 10-80

CPRA 80-100

A

AB

B

O

0-5 yr

5-10 yr

10-15 yr

>15 yr

african american

hispanic

caucasianother

0 20 40 60

Figure 4: Simulated average percentage distributions of recipients across different race, dialysis time,blood type, sensitization, diagnosis and age patient groups for the policy designed in Section 4.5, i.e.,in case study 3 (yellow). The target distributions are also depicted (blue). The results are for anout-of-sample period of 6 months in 2008.

31


33/47

but introduce a change in the required age percentage distribution of recipients. Our method

successfully redesigns a conforming policy.

The change in the age distribution we consider is motivated by comments made by a UNOS

Ethics Committee in OPTNKTC (2008). Based on the fact that, in comparison with current

practice, the new KTC policy would direct a higher number of organs, or organs of higher quality, to

younger patients, the committee argued that the KTC policy might have an unintended consequence

of a decrease in living donor transplants for younger patients, who typically have higher LYFT

scores. In response to that, in this case study we design a policy that has a more balanced age

distribution, which actually resembles the age distribution of patients in the waitlist. Perhaps this

consideration is just only one plausible way of addressing the concern raised by the committee.

Nevertheless, it is presented here only to illustrate the flexibility of our method, rather than tackle

this particular issue.

Finally, note that in comparison with the KTC policy our policy allocates more organs to elder

patients, a fact that could significantly undermine efficiency. However, both policies deliver almost

identical life year gains (see Table 2), which again illustrates that our method is capable of designing

efficient policies.

4.6. Case Study 4

In our final case study we demonstrate how our method can be used to perform a sensitivity analysis

with respect to imposed fairness constraints. Specifically, we explore the dependence of life years

from transplant gains on the priority given for dialysis time and sensitization.

To this end, we consider a similar setup as in Case Study 2, but we relax the constraints

pertaining to patient groups of different dialysis time, i.e., constraints (3), as well as to patient

groups of different sensitization level. The relaxation is controlled by a slack parameter. We then

study the dependence of life year gains on that parameter.

Score components. As in Case Study 2, we use LYFT, a piece-wise linear function of DT (with

break points at 5 and 10 years), CPRA and a step-wise function of patient age (with break points

at 50 and 65 years).

Fairness constraints. We use the same fairness properties with the dominant KTC policy, but

we first relax only the constraints pertaining to dialysis time. The relaxation is performed by

32


34/47

introducing a slack parameter s in the percentage requirements of recipients of different groups,

that is, the relaxed constraints take the form

p: 0DT(p)5

o:(p,o)C

x(p,o) 54 s

100

(p,o)C

x(p,o),

p: 5DT(p)10

o:(p,o)C

x(p,o) 29.5 s

100

(p,o)C

x(p,o),

p: 10DT(p)15

o:(p,o)C

x(p,o) 11.1 s

100

(p,o)C

x(p,o),

p:DT(p)15

o:(p,o)C

x(p,o) 5.4 s

100

(p,o)C

x(p,o).

(4)

Clearly, for s = 0 one would recover the policy that was designed in Case Study 2. For s > 0,

the requirement on matching the percentage distribution (with regard to patient groups of different

dialysis time) achieved by the KTC policy is relaxed. Thus one should expect that policies designed

with such relaxed requirements would achieve higher life years from transplant gains. Using our

method, we design policies for various values of the slack parameter s and quantify how the gains

in medical efficiency depend on deviations from the selected fairness constraints. Secondly, we

follow the same procedure to examine the dependence of medical efficiency on the priority given to

sensitized patients. We again use all the constraints as in Case Study 2, but this time relax only

the constraints pertaining to patient groups of different sensitization levels. The relaxation is again

performed using a slack parameter s. Note that one can potentially perform a sensitivity analysis

through many other different ways of relaxing the constraints; for illustration purposes we focus

here only on the described method of uniformly relaxing the constraints by a slack parameter.

Results. The results we obtain in the aforementioned scenarios are depicted in Figure 5. The

figure shows the life years from transplant gains (for the 6 month period we consider) of policies

designed with relaxed constraints on patient groups of different dialysis time or sensitization, for

various values of the slack parameter s. The figures also depicts the operational point of the KTC

policy, that is for s = 0.

Discussion. Comparing the two scenarios we considered, one can observe that the dependence of

medical efficiency is stronger on dialysis time. Also, the life years from transplant gains can be as

high as 44, 300 years, which are 30% larger than the gains of the KTC policy. Note that although

such a policy might not be implementable, the analysis can provide insights to policymakers and

33


35/47

slack parameter s (in %)

lifeyearsfrom

transplantgains(inthousandsofyears)

50

0

10

10 15

20

20 25

30

40

50

Figure 5: Simulated life years from transplant gains for policies (designed by our method) with relaxedconstraints on all patient groups of different dialysis time (solid) or sensitization (dashed), for variousvalues of the slack parameter s; for more details see Section 4.6. The results are for an out-of-sampleperiod of 6 months in 2008. The marker corresponds to the operational point of the policy proposed bythe UNOS policymakers.

34


36/47

facilitate their decision process.

Nevertheless, this case study illustrates how our method can be used to perform a trade-off

analysis that could assist policymakers in quantifying the impact of certain fairness requirements.

5. Discussion and Future Directions

We dealt with the important problem of allocating deceased donor kidneys to waitlisted patients,

in a fair and efficient way. We focused on national allocation policies in the United States and the

recent effort to revise the current policy in place.

Particularly, we studied allocation policies that are based on point systems; under those policies

patients are awarded points according to some priority criteria, and patients are then prioritized by

the number of points awarded. We identified the challenges in designing a point system, specifically

the relative emphasis put on each criterion such that the resulting policy strikes the right balance

between efficiency and fairness.

Our main contribution was a scalable, data-driven method of designing point system based

allocation policies in an efficient and systematic way. The method does not presume any particular

fairness scheme, or priority criterion. Instead, it offers the flexibility to the designer to select his

desired fairness constraints and criteria under which patients are awarded points. Our method then

balances the criteria and extracts a near-optimal point system policy, in the sense that the policy

outcomes yield approximately the maximum number of life years gains (medical efficiency), while

satisfying the fairness constraints.

To validate our method and demonstrate its usefulness, we presented 4 case studies in which

we designed new policies under different scenarios. In one of them, we designed a new policy

that matches in fairness properties the one that was recently proposed by the U.S. policymakers,

while being based on a format and criteria already considered by policymakers. Critically, our

policy delivers an 8% relative increase in life years gains. The performance gain was established

via simulation, utilizing the same statistical tools and data as the U.S. policymakers.

Finally, we presented a trade-off analysis that revealed the dependence of medical efficiency on

the important fairness concepts of prioritizing patients who have either spent a lot of time waiting,

or are medically incompatible with the majority of donors.

As a pointer for future work, consider the policies that OPTN policymakers have proposed in

35


37/47

which patients and/or organs are categorized into different groups according to some criteria and

then specific groups receive priority in the allocation, see Section 2.1. For instance, a proposal

presented in OPTNKTC (2007) suggests to categorize patients in 5 different groups according to

their expected life year gains: top 20% goes to the first group, bottom 20% goes to the last group,

etc. Similarly, organs are categorized according to their quality (DPI). In the allocation then,

group 1 patients are given priority for group 1 organs, groups 2 patients are given priority for group

2 organs and so on. Ranking within those groups is again achieved via a scoring rule, so our model

would again be applicable and useful. Another interesting question however is, how can should

one decide on the right categorization? In the example we gave, how does one exactly partition

patients into those 5 groups? As an extension and future work, one can potentially use modified

versions of our framework to guide such decisions. We present a related case study in the Appendix.

List of Acronyms

CPRA Calculated Panel Reactive Antibody

DPI Donor Profile Index

DT Dialysis Time

ESRD End-Stage Renal Disease

KAS Kidney Allocation ScoreKPSAM Kidney-Pancreas Simulated Allocation Model

KTC Kidney Transplantation Committee

LYFT Life Years From Transplant

NOTA National Organ Transplant Act

OPO Organ Procurement Organization

OPTN Organ Procurement and Transplantation Network

RFI Request For InformationSRTR Scientific Registry of Transplant Recipients

UNOS United Network for Organ Sharing

36


38/47

Disclaimer

The data reported here have been supplied by the Arbor Research Collaborative for Health (Arbor

Research) as the contractor for the Scientific Registry of Transplant Recipients (SRTR). The inter-

pretation and reporting of these data are the responsibility of the authors and in no way should be

seen as an official policy of or interpretation by the SRTR or the U.S. Government.

References

J. Ahn and J. C. Hornberger. Involving patients in the cadaveric kidney transplant allocation process: A

decision-theoretic perspective. Management Science, 42(5):pp. 629641, 1996.

M. Akan, O. Alagoz, B. Ata, and F. S. Erenay. Optimizing liver allocation system incorporating disease

evolution. Submitted to Operations Research, 2011.

O. Alagoz. Optimal policies for the acceptance of living and cadaveric-donor livers. Ph.D. dissertation,

University of Pittsburgh, Pittsburgh, 2004.

O. Alagoz, L. M. Maillart, A. J. Schaefer, and M. S. Roberts. Determining the acceptance of cadaveric livers

using an implicit model of the waiting list. Operations Research, 55(1):2436, 2007.

I. Ashlagi, D. S. Gilchrist, A. E. Roth, and M. A. Rees. Nonsimultaneous chains and dominos in kidney

paired donation revisited. American Journal of Transplantation, 11(5):984994, May 2011.

I. David and U. Yechiali. A time-dependent stopping problem with application to live organ transplants.Operations Research, 33(3):pp. 491504, 1985.

I. David and U. Yechiali. One-attribute sequential assignment match processes in discrete time. Operations

Research, 43(5):pp. 879884, 1995.

DHHS. Organ Procurement and Transplantation Network Final Rule. 2000. Electronic Code of Federal

Regulations, Title 42Public Health, Chapter IPublic Health Service, Department of Health and

Human Services, Subchapter KHealth Resources Development, Part 121.

D. H. Howard. Why do transplant surgeons turn down organs?: A model of the accept/reject decision.

Journal of Health Economics, 21(6):957 969, 2002.

N. Kong, A. J. Schaefer, B. Hunsaker, and M. S. Roberts. Maximizing the efficiency of the u.s. liver allocation

system through region design. Management Science, 56(12):21112122, 2010. doi: 10.1287/mnsc.1100.

1249.

KPSAM. Kidney-Pancreas Simulated Allocation Model. 2008. Arbor Research Collaborative for Health.

Scientific Registry of Transplant Recipients.

37


39/47


40/47

X. Su and S. A. Zenios. Patient choice in kidney allocation: A sequential stochastic assignment model.

Operations Research, 53(3):443445, May-June 2005.

X. Su and S. A. Zenios. Recipient Choice Can Address the Efficiency-Equity Trade-off in Kidney Transplan-

tation: A Mechanism Design Model. Management Science, 52(11):16471660, 2006.

X. Su and S.A. Zenios. Patient choice in kidney allocation: The role of the queueing discipline. Manufacturing

Service Operations Management, 6(4):280301, 2004. doi: 10.1287/msom.1040.0056.

M. Suthanthiran and T.B. Strom. Renal transplantation. N Engl J Med, page 331:365, 1994.

K. Talluri and G. van Ryzin. An analysis of bid-price controls for network revenue management. Management

Science, 44(11):15771593, November 1998.

UNOS. United Network for Organ Sharing. 2010. http://www.unos.org/.

USRDS. U.S. Renal data system, annual data report: Atlas of chronic kidney disease and end-stage renal

disease in the united states. 2009. National Institutes of Health, National Institute of Diabetes and

Digestive and Kidney Diseases, Bethesda, MD.

L. Waisanen, R.A. Wolfe, R.M. Merion, K. McCullough, and A. Rodgers. Simulating the allocation of organs

for transplantation. Health Care Management Science, 7(4):331338, 2004.

R.A. Wolfe, K.P. McCullough, D.E. Schaubel, J.D. Kalbfleisch, S. Murray, M.D. Stegall, and A.B. Leichtman.

2007 SRTR report on the state of transplantation: Calculating life years from transplant (LYFT):

Methods for kidney and kidney-pancreas candidates. Am J Transplant, 8(2):9971011, 2008.

R.A. Wolfe, K.P. McCullough, and A.B. Leichtman. Predictability of survival models for waiting list and

transplant patients: Calculating LYFT. Am J Transplant, 9(7):15231527, 2009.

S. A. Zenios. Optimal control of a paired-kidney exchange program. Management Science, 48(3):328342,

2002.

S. A. Zenios. Models for kidney allocation. In Margaret Brandeau, Franois Sainfort, and William Pierskalla,

editors, Operations Research and Health Care, volume 70 ofInternational Series in Operations Research

and Management Science, pages 537554. Springer New York, 2005.

S. A. Zenios, G. M. Chertow, and L. M. Wein. Dynamic allocation of kidneys to candidates on the transplant

waiting list. Operations Research, 48(4):pp. 549569, 2000.

39
http://www.unos.org/http://www.unos.org/http://www.unos.org/


41/47

Appendix

A. Simulation Results

Table 3 includes the simulated average percentage distributions (and sample deviations) of recipients

across different race, dialysis time, blood type, sensitization, diagnosis and age patient groups for

the policies designed in the case studies in Section 4, alongside the KTC proposed policy. The

results are for an out-of-sample period of 6 months in 2008.

B. Category-based Policies

As discussed in Section 5, OPTN policymakers have proposed policies in which patients and/or

organs are categorized into different groups according to some criteria. Then, specific groups

receive priority in the allocation of kidneys. For instance, a proposal presented in OPTNKTC

(2007) suggests to categorize patients in 5 different groups according to their expected life year

gains (LYFT): top 20% goes to the first group, bottom 20% goes to the last group, etc. Similarly,

organs are categorized according to their quality (DPI). In the allocation then, group 1 patients

are given priority for group 1 organs, groups 2 patients are given priority for group 2 organs and

so on. In this section we illustrate how one can use a modified version of the framework developed

in Section 3 to decide how one should partition patients and/or organs to such groups.

Suppose we want to design a policy that categorizes patients and organs into groups A and

B, according to their (average) LYFT and DPI scores. Specifically, the top q% of patients are

categorized into group A and the remaining into group B, where q is a policy parameter that needs

to be determined. Similarly we categorize the top q% of organs into group A and the remaining

into group B. Then, once an organ of group A is procured, patients from group A have priority

over group B patients. Patients are then ranked by the time they have spent on dialysis (DT).

If q = 0, the allocation policy is essentially a first-come first-serve policy (FCFS), with respect

to dialysis time. As q increases and we depart from a pure FCFS policy, more emphasis is given

to efficiency as high quality organs are first offered to patients who are likely to benefit more from

them (in terms of life year gains). Thus, one expects higher LYFT gains. As q approaches 100

however, we again recover FCFS. Below we discuss a way of guiding the selection of q based on our

40


42/47


43/47

framework.

For the case of FCFS, i.e., q = 0 or q = 100, one can solve problem (1) without any fairness

constraints and with objective coefficients equal to DT(p) for every (p,o) pair, to obtain an ap-

proximate expected allocation. Suppose we now consider 0 < q < 100. Let AP(q) and AO(q) be

the sets of all group A patients and organs respectively. We then define an adjusted set of eligible

patient-organ pairs C(q) (see Section 3.1), in which group A organs are only allocated to group A

patients (according to the selected value of q), i.e.,

C(q) = {(p,o) : (p,o) C and p AP(q) if o AO(q)} .

One can then solve the following problem to obtain an approximate expected allocation for any

selection of q:

maximize

(p,o)C(q)

DT(p)x(p,o)

subject to

o:(p,o)C(q)

x(p,o) 1, p

p:(p,o)C(q)

x(p,o) 1, o

x 0.

(5)

Let x(q) denote an optimal solution of (5) for any selection of q.

To quantitavely characterize the efficiency of each of those policies (as we vary q), one can

calcu

bertsimas 2011) fairness-efficiency kidney transplantation

Documents